Skip to content
  1. Apr 29, 2008
  3. Apr 28, 2008
  5. Apr 27, 2008
    • Chris Lattner's avatar
      Implement a signficant optimization for inline asm: · 22379734
      Chris Lattner authored 
      When choosing between constraints with multiple options,
like "ir", test to see if we can use the 'i' constraint and
go with that if possible.  This produces more optimal ASM in
all cases (sparing a register and an instruction to load it),
and fixes inline asm like this:

void test () {
  asm volatile (" %c0 %1 " : : "imr" (42), "imr"(14));
}

Previously we would dump "42" into a memory location (which
is ok for the 'm' constraint) which would cause a problem
because the 'c' modifier is not valid on memory operands.

Isn't it great how inline asm turns 'missed optimization'
into 'compile failed'??

Incidentally, this was the todo in 
PowerPC/2007-04-24-InlineAsm-I-Modifier.ll

Please do NOT pull this into Tak.

llvm-svn: 50315
      22379734
    • Chris Lattner's avatar
      Move a bunch of inline asm code out of line. · 4793515a
      Chris Lattner authored 
      llvm-svn: 50313
      4793515a
  7. Apr 26, 2008
  9. Apr 25, 2008
  11. Apr 24, 2008
  13. Apr 23, 2008
  15. Apr 22, 2008
    • Chris Lattner's avatar
      Start doing the significantly useful part of jump threading: handle cases · 37e9c187
      Chris Lattner authored 
      where a comparison has a phi input and that phi is a constant.  For example,
stuff like:

  Threading edge through bool from 'bb2149' to 'bb2231' with cost: 1, across block:
bb2237:		; preds = %bb2231, %bb2149
	%tmp2328.rle = phi i32 [ %tmp2232, %bb2231 ], [ %tmp2232439, %bb2149 ]		; <i32> [#uses=2]
	%done.0 = phi i32 [ %done.2, %bb2231 ], [ 0, %bb2149 ]		; <i32> [#uses=1]
	%tmp2239 = icmp eq i32 %done.0, 0		; <i1> [#uses=1]
	br i1 %tmp2239, label %bb2231, label %bb2327

or

bb38.i298:		; preds = %bb33.i295, %bb1693
	%tmp39.i296.rle = phi %struct.ibox* [ null, %bb1693 ], [ %tmp39.i296.rle1109, %bb33.i295 ]		; <%struct.ibox*> [#uses=2]
	%minspan.1.i291.reg2mem.1 = phi i32 [ 32000, %bb1693 ], [ %minspan.0.i288, %bb33.i295 ]		; <i32> [#uses=1]
	%tmp40.i297 = icmp eq %struct.ibox* %tmp39.i296.rle, null		; <i1> [#uses=1]
	br i1 %tmp40.i297, label %implfeeds.exit311, label %bb43.i301

This triggers thousands of times in spec.

llvm-svn: 50110
      37e9c187
    • Chris Lattner's avatar
      Dig through multiple levels of AND to thread jumps if needed. · d5425e8f
      Chris Lattner authored 
      llvm-svn: 50106
      d5425e8f
    • Chris Lattner's avatar
      Teach jump threading to thread through blocks like: · 3df4c15d
      Chris Lattner authored 
        br (and X, phi(Y, Z, false)), label L1, label L2

This triggers once on 252.eon and 6 times on 176.gcc.  Blocks 
in question often look like this:

bb262:		; preds = %bb261, %bb248
	%iftmp.251.0 = phi i1 [ true, %bb261 ], [ false, %bb248 ]		; <i1> [#uses=4]
	%tmp270 = icmp eq %struct.rtx_def* %tmp.0.i, null		; <i1> [#uses=1]
	%bothcond = or i1 %iftmp.251.0, %tmp270		; <i1> [#uses=1]
	br i1 %bothcond, label %bb288, label %bb273

In this case, it is clear that it doesn't matter if tmp.0.i is null when coming from bb261.  When coming from bb248, it is all that matters.


Another random example:

check_asm_operands.exit:		; preds = %check_asm_operands.exit.thr_comm, %bb30.i, %bb12.i, %bb6.i413
	%tmp.0.i420 = phi i1 [ true, %bb6.i413 ], [ true, %bb12.i ], [ true, %bb30.i ], [ false, %check_asm_operands.exit.thr_comm ; <i1> [#uses=1]
	call void @llvm.stackrestore( i8* %savedstack ) nounwind 
	%tmp4389 = icmp eq i32 %added_sets_1.0, 0		; <i1> [#uses=1]
	%tmp4394 = icmp eq i32 %added_sets_2.0, 0		; <i1> [#uses=1]
	%bothcond80 = and i1 %tmp4389, %tmp4394		; <i1> [#uses=1]
	%bothcond81 = and i1 %bothcond80, %tmp.0.i420		; <i1> [#uses=1]
	br i1 %bothcond81, label %bb4398, label %bb4397

Here is the case from 252.eon:

bb290.i.i:		; preds = %bb23.i57.i.i, %bb8.i39.i.i, %bb100.i.i, %bb100.i.i, %bb85.i.i110
	%myEOF.1.i.i = phi i1 [ true, %bb100.i.i ], [ true, %bb100.i.i ], [ true, %bb85.i.i110 ], [ true, %bb8.i39.i.i ], [ false, %bb23.i57.i.i ]		; <i1> [#uses=2]
	%i.4.i.i = phi i32 [ %i.1.i.i, %bb85.i.i110 ], [ %i.0.i.i, %bb100.i.i ], [ %i.0.i.i, %bb100.i.i ], [ %i.3.i.i, %bb8.i39.i.i ], [ %i.3.i.i, %bb23.i57.i.i ]		; <i32> [#uses=3]
	%tmp292.i.i = load i8* %tmp16.i.i100, align 1		; <i8> [#uses=1]
	%tmp293.not.i.i = icmp ne i8 %tmp292.i.i, 0		; <i1> [#uses=1]
	%bothcond.i.i = and i1 %tmp293.not.i.i, %myEOF.1.i.i		; <i1> [#uses=1]
	br i1 %bothcond.i.i, label %bb202.i.i, label %bb301.i.i
  Factoring out 3 common predecessors.

On the path from any blocks other than bb23.i57.i.i, the load and compare 
are dead.

llvm-svn: 50096
      3df4c15d
    • Chris Lattner's avatar
      refactor some code, no functionality change. · e369c35a
      Chris Lattner authored 
      llvm-svn: 50094
      e369c35a
    • Chris Lattner's avatar
      remove dead code. · 8fb13cbe
      Chris Lattner authored 
      llvm-svn: 50080
      8fb13cbe
    • Chris Lattner's avatar
      optimize "p != gep p, ..." better. This allows us to compile · c3a43935
      Chris Lattner authored 
      getelementptr-seteq.ll into:

define i1 @test(i64 %X, %S* %P) {
	%C = icmp eq i64 %X, -1		; <i1> [#uses=1]
	ret i1 %C
}

instead of:

define i1 @test(i64 %X, %S* %P) {
	%A.idx.mask = and i64 %X, 4611686018427387903		; <i64> [#uses=1]
	%C = icmp eq i64 %A.idx.mask, 4611686018427387903		; <i1> [#uses=1]
	ret i1 %C
}

And fixes the second half of PR2235.  This speeds up the insertion sort
case by 45%, from 1.12s to 0.77s.  In practice, this will significantly
speed up for loops structured like:

for (double *P = Base + N; P != Base; --P)
  ...

Which happens frequently for C++ iterators.

llvm-svn: 50079
      c3a43935
  17. Apr 21, 2008